1067
Killswitch Engineer (lemmy.world)
submitted 10 months ago by to c/programmer_humor@programming.dev
77 comments fedilink hide all child comments
 
you are viewing a single comment's thread
view the rest of the comments
[–] 10 points 10 months ago* (last edited 10 months ago) (4 children)

Everyone here so far has forgotten that in simulations, the model has blackmailed the person responsible shutting it off and even gone so far as to cancel active alerts in order to prevent an executive laying unconscous in the server room from receiving life-saving care.

  • permalink
  • fedilink
  • source
  • hideshow 4 child comments
    •
  • [–] 13 points 10 months ago* (last edited 10 months ago) (3 children)

    The model 'blackmailed' the person because they provided it with a prompt asking it to pretend to blackmail them. Gee, I wonder what they expected.

    Have not heard the one about cancelling active alerts, but I doubt it's any less bullshit. Got a source about it?

    Edit: Here's a deep dive into why those claims are BS: https://www.aipanic.news/p/ai-blackmail-fact-checking-a-misleading

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] 3 points 10 months ago (2 children)

    I provided enough information that the relevant source shows up in a search, but here you go:

    In no situation did we explicitly instruct any models to blackmail or do any of the other harmful actions we observe. [Lynch, et al., "Agentic Misalignment: How LLMs Could be an Insider Threat", Anthropic Research, 2025]

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •